Spaces:
Running
Running
Sadjad Alikhani
commited on
Update app.py
Browse files
app.py
CHANGED
@@ -50,72 +50,50 @@ def display_predefined_images(percentage_idx):
|
|
50 |
|
51 |
return raw_image, embeddings_image
|
52 |
|
53 |
-
# Updated los_nlos_classification to handle missing outputs properly
|
54 |
-
def los_nlos_classification(file, percentage_idx):
|
55 |
-
if file is not None:
|
56 |
-
raw_cm_image, emb_cm_image, console_output = process_hdf5_file(file, percentage_idx)
|
57 |
-
return raw_cm_image, emb_cm_image, console_output
|
58 |
-
else:
|
59 |
-
raw_image, embeddings_image = display_predefined_images(percentage_idx)
|
60 |
-
return raw_image, embeddings_image, "No file uploaded. Displaying predefined images."
|
61 |
-
|
62 |
# Function to create random images for LoS/NLoS classification results
|
63 |
def create_random_image(size=(300, 300)):
|
64 |
random_image = np.random.rand(*size, 3) * 255
|
65 |
return Image.fromarray(random_image.astype('uint8'))
|
66 |
|
67 |
-
# Function to load the pre-trained model from your cloned repository
|
68 |
-
def load_custom_model():
|
69 |
-
from lwm_model import LWM # Assuming the model is defined in lwm_model.py
|
70 |
-
model = LWM() # Modify this according to your model initialization
|
71 |
-
model.eval()
|
72 |
-
return model
|
73 |
-
|
74 |
-
import importlib.util
|
75 |
-
|
76 |
-
# Function to dynamically load a Python module from a given file path
|
77 |
-
def load_module_from_path(module_name, file_path):
|
78 |
-
spec = importlib.util.spec_from_file_location(module_name, file_path)
|
79 |
-
module = importlib.util.module_from_spec(spec)
|
80 |
-
spec.loader.exec_module(module)
|
81 |
-
return module
|
82 |
-
|
83 |
# Function to split dataset into training and test sets based on user selection
|
84 |
-
def
|
85 |
-
|
86 |
-
num_samples = channels.shape[0]
|
87 |
-
train_size = int(num_samples * percentage)
|
88 |
-
print(f'Number of Training Samples: {train_size}')
|
89 |
|
90 |
-
indices
|
91 |
-
|
92 |
|
93 |
-
|
|
|
|
|
94 |
|
95 |
-
|
96 |
-
|
|
|
97 |
|
98 |
-
|
|
|
|
|
|
|
|
|
|
|
99 |
|
100 |
-
|
101 |
-
|
102 |
-
return np.linalg.norm(x - centroid)
|
103 |
|
104 |
-
|
105 |
|
|
|
106 |
def classify_based_on_distance(train_data, train_labels, test_data):
|
107 |
-
|
108 |
-
|
109 |
-
centroid_1 = train_data[train_labels == 1].mean(dim=0) # Use torch.mean
|
110 |
|
111 |
predictions = []
|
112 |
for test_point in test_data:
|
113 |
-
|
114 |
-
|
115 |
-
dist_1 = euclidean_distance(test_point, centroid_1)
|
116 |
predictions.append(0 if dist_0 < dist_1 else 1)
|
117 |
|
118 |
-
return torch.tensor(predictions)
|
119 |
|
120 |
# Function to generate confusion matrix plot
|
121 |
def plot_confusion_matrix(y_true, y_pred, title):
|
@@ -132,121 +110,75 @@ def plot_confusion_matrix(y_true, y_pred, title):
|
|
132 |
plt.savefig(f"{title}.png")
|
133 |
return Image.open(f"{title}.png")
|
134 |
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
# Generate the indices for shuffling and splitting
|
139 |
-
indices = torch.randperm(N) # Randomly shuffle the indices
|
140 |
-
|
141 |
-
# Calculate the split index
|
142 |
-
split_index = int(N * percentage_values[percentage_idx])
|
143 |
-
print(f'Training Size: {split_index}')
|
144 |
-
|
145 |
-
# Split indices into train and test
|
146 |
-
train_indices = indices[:split_index] # First 80% for training
|
147 |
-
test_indices = indices[split_index:] # Remaining 20% for testing
|
148 |
-
|
149 |
-
# Select the same indices from both output_emb and output_raw
|
150 |
-
train_emb = output_emb[train_indices]
|
151 |
-
test_emb = output_emb[test_indices]
|
152 |
-
|
153 |
-
train_raw = output_raw[train_indices]
|
154 |
-
test_raw = output_raw[test_indices]
|
155 |
-
|
156 |
-
train_labels = labels[train_indices]
|
157 |
-
test_labels = labels[test_indices]
|
158 |
-
|
159 |
-
return train_emb, test_emb, train_raw, test_raw, train_labels, test_labels
|
160 |
-
|
161 |
-
# Store the original working directory when the app starts
|
162 |
-
original_dir = os.getcwd()
|
163 |
-
|
164 |
-
def process_hdf5_file(uploaded_file, percentage_idx):
|
165 |
capture = PrintCapture()
|
166 |
sys.stdout = capture # Redirect print statements to capture
|
167 |
-
|
168 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
169 |
model_repo_url = "https://huggingface.co/sadjadalikhani/LWM"
|
170 |
model_repo_dir = "./LWM"
|
171 |
|
172 |
-
# Step 1: Clone the repository if not already done
|
173 |
if not os.path.exists(model_repo_dir):
|
174 |
print(f"Cloning model repository from {model_repo_url}...")
|
175 |
subprocess.run(["git", "clone", model_repo_url, model_repo_dir], check=True)
|
176 |
|
177 |
-
#
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
print(f"Changed working directory to {os.getcwd()}")
|
182 |
-
print(f"Directory content: {os.listdir(os.getcwd())}") # Debugging: Check repo content
|
183 |
-
else:
|
184 |
-
print(f"Directory {repo_work_dir} does not exist.")
|
185 |
-
return
|
186 |
-
|
187 |
-
# Step 3: Dynamically load lwm_model.py, input_preprocess.py, and inference.py
|
188 |
-
lwm_model_path = os.path.join(os.getcwd(), 'lwm_model.py')
|
189 |
-
input_preprocess_path = os.path.join(os.getcwd(), 'input_preprocess.py')
|
190 |
-
inference_path = os.path.join(os.getcwd(), 'inference.py')
|
191 |
-
|
192 |
-
# Load lwm_model
|
193 |
-
lwm_model = load_module_from_path("lwm_model", lwm_model_path)
|
194 |
|
195 |
-
# Load
|
|
|
196 |
input_preprocess = load_module_from_path("input_preprocess", input_preprocess_path)
|
197 |
-
|
198 |
-
# Load inference
|
199 |
inference = load_module_from_path("inference", inference_path)
|
200 |
|
201 |
-
# Step 4: Load the model from lwm_model module
|
202 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
203 |
-
print(f"Loading
|
204 |
model = lwm_model.LWM.from_pretrained(device=device).to(torch.float32)
|
205 |
|
206 |
-
#
|
207 |
-
with h5py.File(uploaded_file.name, 'r') as f:
|
208 |
-
channels = np.array(f['channels']) # Assuming 'channels' dataset in the HDF5 file
|
209 |
-
labels = np.array(f['labels']) # Assuming 'labels' dataset in the HDF5 file
|
210 |
-
print(f"Loaded dataset with {channels.shape[0]} samples.")
|
211 |
-
|
212 |
-
# Step 7: Tokenize the data using the tokenizer from input_preprocess
|
213 |
preprocessed_chs = input_preprocess.tokenizer(manual_data=channels)
|
214 |
-
print(preprocessed_chs[0][0][1])
|
215 |
-
|
216 |
-
# Step 7: Perform inference using the functions from inference.py
|
217 |
output_emb = inference.lwm_inference(preprocessed_chs, 'channel_emb', model)
|
218 |
-
#print(f'output_emb:{output_emb[10][0]}')
|
219 |
output_raw = inference.create_raw_dataset(preprocessed_chs, device)
|
220 |
-
#print(f'output_raw:{output_raw[10][0]}')
|
221 |
|
222 |
print(f"Output Embeddings Shape: {output_emb.shape}")
|
223 |
print(f"Output Raw Shape: {output_raw.shape}")
|
224 |
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
229 |
|
230 |
-
# Step 8: Perform classification using the Euclidean distance for both raw and embeddings
|
231 |
-
print(f'train_data_emb: {train_data_emb.shape}')
|
232 |
-
print(f'train_labels: {train_labels.shape}')
|
233 |
-
print(f'test_data_emb: {test_data_emb.shape}')
|
234 |
pred_raw = classify_based_on_distance(train_data_raw, train_labels, test_data_raw)
|
235 |
pred_emb = classify_based_on_distance(train_data_emb, train_labels, test_data_emb)
|
236 |
-
|
237 |
-
# Step 9: Generate confusion matrices for both raw and embeddings
|
238 |
raw_cm_image = plot_confusion_matrix(test_labels, pred_raw, title="Confusion Matrix (Raw Channels)")
|
239 |
emb_cm_image = plot_confusion_matrix(test_labels, pred_emb, title="Confusion Matrix (Embeddings)")
|
240 |
|
241 |
-
return raw_cm_image, emb_cm_image,
|
242 |
-
|
243 |
-
|
244 |
-
return str(e), str(e), capture.get_output()
|
245 |
-
|
246 |
-
finally:
|
247 |
-
# Always return to the original working directory after processing
|
248 |
-
os.chdir(original_dir)
|
249 |
-
sys.stdout = sys.__stdout__ # Reset print statements
|
250 |
|
251 |
# Define the Gradio interface
|
252 |
with gr.Blocks(css="""
|
@@ -262,56 +194,37 @@ with gr.Blocks(css="""
|
|
262 |
text-align: center;
|
263 |
}
|
264 |
""") as demo:
|
265 |
-
|
266 |
-
# Contact Section
|
267 |
-
gr.Markdown("""
|
268 |
-
<div style="text-align: center;">
|
269 |
-
<a target="_blank" href="https://www.wi-lab.net">
|
270 |
-
<img src="https://www.wi-lab.net/wp-content/uploads/2021/08/WI-name.png" alt="Wireless Model" style="height: 30px;">
|
271 |
-
</a>
|
272 |
-
<a target="_blank" href="mailto:alikhani@asu.edu" style="margin-left: 10px;">
|
273 |
-
<img src="https://img.shields.io/badge/email-alikhani@asu.edu-blue.svg?logo=gmail" alt="Email">
|
274 |
-
</a>
|
275 |
-
</div>
|
276 |
-
""")
|
277 |
-
|
278 |
-
# Tabs for Beam Prediction and LoS/NLoS Classification
|
279 |
-
with gr.Tab("Beam Prediction Task"):
|
280 |
-
gr.Markdown("### Beam Prediction Task")
|
281 |
-
|
282 |
-
with gr.Row():
|
283 |
-
with gr.Column(elem_id="slider-container"):
|
284 |
-
gr.Markdown("Percentage of Data for Training")
|
285 |
-
percentage_slider_bp = gr.Slider(minimum=0, maximum=4, step=1, value=0, interactive=True, elem_id="vertical-slider")
|
286 |
-
|
287 |
-
with gr.Row():
|
288 |
-
raw_img_bp = gr.Image(label="Raw Channels", type="pil", width=300, height=300, interactive=False)
|
289 |
-
embeddings_img_bp = gr.Image(label="Embeddings", type="pil", width=300, height=300, interactive=False)
|
290 |
-
|
291 |
-
percentage_slider_bp.change(fn=display_predefined_images, inputs=[percentage_slider_bp], outputs=[raw_img_bp, embeddings_img_bp])
|
292 |
|
|
|
293 |
with gr.Tab("LoS/NLoS Classification Task"):
|
294 |
gr.Markdown("### LoS/NLoS Classification Task")
|
295 |
-
|
296 |
file_input = gr.File(label="Upload HDF5 Dataset", file_types=[".h5"])
|
297 |
|
298 |
with gr.Row():
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
interactive=True)
|
306 |
|
307 |
with gr.Row():
|
308 |
raw_img_los = gr.Image(label="Raw Channels", type="pil", width=300, height=300, interactive=False)
|
309 |
embeddings_img_los = gr.Image(label="Embeddings", type="pil", width=300, height=300, interactive=False)
|
310 |
output_textbox = gr.Textbox(label="Console Output", lines=10)
|
311 |
|
312 |
-
|
313 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
314 |
|
315 |
# Launch the app
|
316 |
if __name__ == "__main__":
|
317 |
-
demo.launch()
|
|
|
50 |
|
51 |
return raw_image, embeddings_image
|
52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
# Function to create random images for LoS/NLoS classification results
|
54 |
def create_random_image(size=(300, 300)):
|
55 |
random_image = np.random.rand(*size, 3) * 255
|
56 |
return Image.fromarray(random_image.astype('uint8'))
|
57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
# Function to split dataset into training and test sets based on user selection
|
59 |
+
def identical_train_test_split(output_emb, output_raw, labels, percentage_idx):
|
60 |
+
N = output_emb.shape[0] # Get the total number of samples
|
|
|
|
|
|
|
61 |
|
62 |
+
# Generate the indices for shuffling and splitting
|
63 |
+
indices = torch.randperm(N) # Randomly shuffle the indices
|
64 |
|
65 |
+
# Calculate the split index
|
66 |
+
split_index = int(N * percentage_values[percentage_idx] / 10) # Convert percentage index to percentage value
|
67 |
+
print(f'Training Size: {split_index}')
|
68 |
|
69 |
+
# Split indices into train and test
|
70 |
+
train_indices = indices[:split_index]
|
71 |
+
test_indices = indices[split_index:]
|
72 |
|
73 |
+
# Select the same indices from both output_emb and output_raw
|
74 |
+
train_emb = output_emb[train_indices]
|
75 |
+
test_emb = output_emb[test_indices]
|
76 |
+
|
77 |
+
train_raw = output_raw[train_indices]
|
78 |
+
test_raw = output_raw[test_indices]
|
79 |
|
80 |
+
train_labels = labels[train_indices]
|
81 |
+
test_labels = labels[test_indices]
|
|
|
82 |
|
83 |
+
return train_emb, test_emb, train_raw, test_raw, train_labels, test_labels
|
84 |
|
85 |
+
# Function to calculate Euclidean distance between a point and a centroid
|
86 |
def classify_based_on_distance(train_data, train_labels, test_data):
|
87 |
+
centroid_0 = train_data[train_labels == 0].mean(dim=0)
|
88 |
+
centroid_1 = train_data[train_labels == 1].mean(dim=0)
|
|
|
89 |
|
90 |
predictions = []
|
91 |
for test_point in test_data:
|
92 |
+
dist_0 = torch.norm(test_point - centroid_0)
|
93 |
+
dist_1 = torch.norm(test_point - centroid_1)
|
|
|
94 |
predictions.append(0 if dist_0 < dist_1 else 1)
|
95 |
|
96 |
+
return torch.tensor(predictions)
|
97 |
|
98 |
# Function to generate confusion matrix plot
|
99 |
def plot_confusion_matrix(y_true, y_pred, title):
|
|
|
110 |
plt.savefig(f"{title}.png")
|
111 |
return Image.open(f"{title}.png")
|
112 |
|
113 |
+
# Function to handle inference and return the results (store the results to state)
|
114 |
+
def run_inference(uploaded_file):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
capture = PrintCapture()
|
116 |
sys.stdout = capture # Redirect print statements to capture
|
117 |
+
|
118 |
try:
|
119 |
+
# Load the HDF5 file and extract channels and labels
|
120 |
+
with h5py.File(uploaded_file.name, 'r') as f:
|
121 |
+
channels = np.array(f['channels']) # Assuming 'channels' dataset in the HDF5 file
|
122 |
+
labels = np.array(f['labels']) # Assuming 'labels' dataset in the HDF5 file
|
123 |
+
print(f"Loaded dataset with {channels.shape[0]} samples.")
|
124 |
+
|
125 |
+
# Run the tokenization and model inference
|
126 |
model_repo_url = "https://huggingface.co/sadjadalikhani/LWM"
|
127 |
model_repo_dir = "./LWM"
|
128 |
|
|
|
129 |
if not os.path.exists(model_repo_dir):
|
130 |
print(f"Cloning model repository from {model_repo_url}...")
|
131 |
subprocess.run(["git", "clone", model_repo_url, model_repo_dir], check=True)
|
132 |
|
133 |
+
# Load the model
|
134 |
+
lwm_model_path = os.path.join(model_repo_dir, 'lwm_model.py')
|
135 |
+
input_preprocess_path = os.path.join(model_repo_dir, 'input_preprocess.py')
|
136 |
+
inference_path = os.path.join(model_repo_dir, 'inference.py')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
137 |
|
138 |
+
# Load dynamically
|
139 |
+
lwm_model = load_module_from_path("lwm_model", lwm_model_path)
|
140 |
input_preprocess = load_module_from_path("input_preprocess", input_preprocess_path)
|
|
|
|
|
141 |
inference = load_module_from_path("inference", inference_path)
|
142 |
|
|
|
143 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
144 |
+
print(f"Loading LWM model on {device}...")
|
145 |
model = lwm_model.LWM.from_pretrained(device=device).to(torch.float32)
|
146 |
|
147 |
+
# Preprocess and inference
|
|
|
|
|
|
|
|
|
|
|
|
|
148 |
preprocessed_chs = input_preprocess.tokenizer(manual_data=channels)
|
|
|
|
|
|
|
149 |
output_emb = inference.lwm_inference(preprocessed_chs, 'channel_emb', model)
|
|
|
150 |
output_raw = inference.create_raw_dataset(preprocessed_chs, device)
|
|
|
151 |
|
152 |
print(f"Output Embeddings Shape: {output_emb.shape}")
|
153 |
print(f"Output Raw Shape: {output_raw.shape}")
|
154 |
|
155 |
+
return output_emb, output_raw, labels, capture.get_output()
|
156 |
+
|
157 |
+
except Exception as e:
|
158 |
+
return None, None, None, str(e)
|
159 |
+
|
160 |
+
finally:
|
161 |
+
sys.stdout = sys.__stdout__ # Reset print statements
|
162 |
+
|
163 |
+
# Function to handle classification after inference (using Gradio state)
|
164 |
+
def los_nlos_classification(output_emb, output_raw, labels, percentage_idx):
|
165 |
+
if output_emb is not None and output_raw is not None:
|
166 |
+
train_data_emb, test_data_emb, train_data_raw, test_data_raw, train_labels, test_labels = identical_train_test_split(
|
167 |
+
output_emb.view(len(output_emb), -1),
|
168 |
+
output_raw.view(len(output_raw), -1),
|
169 |
+
labels,
|
170 |
+
percentage_idx
|
171 |
+
)
|
172 |
|
|
|
|
|
|
|
|
|
173 |
pred_raw = classify_based_on_distance(train_data_raw, train_labels, test_data_raw)
|
174 |
pred_emb = classify_based_on_distance(train_data_emb, train_labels, test_data_emb)
|
175 |
+
|
|
|
176 |
raw_cm_image = plot_confusion_matrix(test_labels, pred_raw, title="Confusion Matrix (Raw Channels)")
|
177 |
emb_cm_image = plot_confusion_matrix(test_labels, pred_emb, title="Confusion Matrix (Embeddings)")
|
178 |
|
179 |
+
return raw_cm_image, emb_cm_image, "Classification successful"
|
180 |
+
|
181 |
+
return create_random_image(), create_random_image(), "No valid inference outputs"
|
|
|
|
|
|
|
|
|
|
|
|
|
182 |
|
183 |
# Define the Gradio interface
|
184 |
with gr.Blocks(css="""
|
|
|
194 |
text-align: center;
|
195 |
}
|
196 |
""") as demo:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
197 |
|
198 |
+
# Tabs for Beam Prediction and LoS/NLoS Classification
|
199 |
with gr.Tab("LoS/NLoS Classification Task"):
|
200 |
gr.Markdown("### LoS/NLoS Classification Task")
|
201 |
+
|
202 |
file_input = gr.File(label="Upload HDF5 Dataset", file_types=[".h5"])
|
203 |
|
204 |
with gr.Row():
|
205 |
+
percentage_dropdown_los = gr.Dropdown(
|
206 |
+
choices=[str(v) for v in percentage_values * 10],
|
207 |
+
value=10,
|
208 |
+
label="Percentage of Data for Training",
|
209 |
+
interactive=True
|
210 |
+
)
|
|
|
211 |
|
212 |
with gr.Row():
|
213 |
raw_img_los = gr.Image(label="Raw Channels", type="pil", width=300, height=300, interactive=False)
|
214 |
embeddings_img_los = gr.Image(label="Embeddings", type="pil", width=300, height=300, interactive=False)
|
215 |
output_textbox = gr.Textbox(label="Console Output", lines=10)
|
216 |
|
217 |
+
# Process file upload to run inference
|
218 |
+
inference_output = gr.State()
|
219 |
+
file_input.upload(run_inference, inputs=file_input, outputs=inference_output)
|
220 |
+
|
221 |
+
# Handle dropdown change for classification
|
222 |
+
percentage_dropdown_los.change(
|
223 |
+
fn=los_nlos_classification,
|
224 |
+
inputs=[inference_output['output_emb'], inference_output['output_raw'], inference_output['labels'], percentage_dropdown_los],
|
225 |
+
outputs=[raw_img_los, embeddings_img_los, output_textbox]
|
226 |
+
)
|
227 |
|
228 |
# Launch the app
|
229 |
if __name__ == "__main__":
|
230 |
+
demo.launch()
|